1. CSS434 Presentation Guide
# slides should be around 15 for a 20-minute talk.
Show a table of contents, (i.e., what you will be talking about).
Get started with the background of the project you surveyed.
Digest the essense of the project rather than cut and paste the contes from the papers you read.
Include examples, illustrations, and performance results.
Clarify pros and cons of the research/development project you surveyed.
Add your opinion to improve the project.
Conclude your presentation.

2. CSS434 Demo Talk Agent-Based Traffic Simulation
Munehiro Fukuda
University of Washington Bothell

3. Table of Contents Conventional Mathematical Models
CSS434 Demo Talk: Agent-Based Traffic Simulation
Table of Contents
Conventional Mathematical Models
Micro-Simulation: Agent-Based Models
MATSim
Challenges in Agent-Based Transport Simulation
Summary

4. Backgound Macroscopic Simulation
CSS434 Demo Talk: Agent-Based Traffic Simulation
Backgound Macroscopic Simulation
Merits
Demerits
Mathematical models
General parameter assumptions
Construction, fires, etc. considered as bias to the model
Ease of real data retrieval such as highway traffic
WSDOT annual traffic report
Mathematical verification
No micro events or interactions considered
Traffic signals and lanes
Parking
Freight traffic
Public transport
No dynamic events considered
Weather
Dynamic trip plans

5. Background Agent-Based Modeling
CSS434 Demo Talk: Agent-Based Traffic Simulation
Background Agent-Based Modeling
Micro-simulation
Views interaction among a large number of simulation entities, (a.k.a. agents).
Simulates an emergent collective group behavior of agents
Agent-based transport simulation
Model each traveler as an agent.
Consider as many traffic events as possible.
Simulates traffic as an interaction among travelers and events.
System examples (open source)
TRANSIMS: based on cellular automata
MATSim: based on a queuing network

6. MATSim Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
MATSim
Variable lenth
Event-based queuing simulation
XML input files
Network configuration
Log File
Score Statistics
Leg Travel Distance Statistics
Events
Trip Durations
Optimization is performed in terms of agents’ plans.
10% agents: reroute their plans dynamically.
90% agents: choose their best score.
Agent plans
From Andreas Horni, Kai Nagel, Kay W. Axhausen, “The Multi-Agent Transport Simulation MATSim”

7. MATSim Example Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
MATSim Example
From

8. Challenges in Agent-Based Models
Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
Challenges in Agent-Based Models
Modeling
A huge manpower would be required to model signals, lanes, parking, etc. in details rather than to give global models and parameters.
Calibration
Non-mathematical verifications are difficult to trust.
How much detailed data can be sampled from the real world?
Computation
Millions of agents drive through several thousands of cells in TRANSIMS and links in MATSim.

9. Modeling in Agent-Based Models
Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
Modeling in Agent-Based Models
Links: speed, signals, and lanes
Parking
Public transport
Freight traffic
Dynamic events (e.g., accidents and weather changes)
Pro: Agents and micro-simulation can describe almost whatever we want to model.

10. Public Transport in MATSim
Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
Public Transport in MATSim
Teleportion
An agent is removed from one location and place at a later point of time.
TransitVehicles.xml
Vehicle type
Passenger capasity
Actual vehicles
TransitSchedule.xml
Transit stops with names
Transit lines
Routes (links) used by the transit
Schedules
Teleportion
Con: Labor/time/data-intensive work
From Andreas Horni, Kai Nagel, Kay W. Axhausen, “The Multi-Agent Transport Simulation MATSim”

11. Calibrations in Agent-Based Models
Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
Calibrations in Agent-Based Models
Realism tests *
Hourly traffic flows: can be compared with automatic traffic recorders’ data
Travel times and speeds: can be compared with public transports’ data
Traffic patterns (queuing patterns at intersections, congested roads, freeway lane choice, merging, etc.): traffic cameras??
Available traffic data *
Automatic traffic recorder(s)’ samples are sparse and imperfect.
Drivers’ mentality (e.g., aggressiveness) varies in metropolitan and suburban areas.
It is impossible to prepare millions of all agent itineraries and perturbations, thus we need to sample householders’ data.
Comparing simulation results with real data
Data-intensive and labor-intensive work *
* From Wisconsin DOT Micro-Simulation Guideline:

12. Computation in Agent-Based Models
CSS434 Demo Talk: Agent-Based Traffic Simulation
Computation in Agent-Based Models
Large # road links
A TRANSIMS simulation of 200,000 links in Portland: 0.23 sec per simulation step (1 sec)
From Kai Nagel, Marcus Ricket, “Parallel implementation of the TRANSIMS micro-simulation”, Parallel Computing Vol 27(N.12), 2001
A day traffic simulation would take 5.5 hours.
Large # agents
A MATSim simulation of 10,000-car circular movement over 10,000 links: 51 sec
From John Piger, MASS library traffic simulation application development and performance evaluation. Css497 final report, University of Washington, Bothell, WA, August 2011
A movement of 200,000 cars driving through I-405 in Bellevue would take 17 minutes, then a day traffic simulation?
Solution: Parallel and distributed simulation

13. Future Distributed Computing in MATSim
CSS434 Demo Talk: Agent-Based Traffic Simulation
Future Distributed Computing in MATSim
Master-slave mode
Qsim on master
Runs selected plans in a full queue simulation.
Uses multithreading for parallelization.
Psim on slave nodes
Produce and evaluate plans for all agents.
Pro
Could distribute agents over a cluster and reduce memory usage per node.
Con
Would still suffer from CPU-intensive micro-simulation.
From Andreas Horni, Kai Nagel, Kay W. Axhausen, “The Multi-Agent Transport Simulation MATSim”

14. Our Approach to MATSim Parallelization
Global City Teams Challenge Super Action Cluster Summit
Feb 2, 2017
Our Approach to MATSim Parallelization
Decentralized model
Qsim on all computing node
Links and nodes are mapped to a distributed array
Agents migrated over a distributed space.
Performance
Pro: Better than the original multithreaded MATSim
Con: Load balancing needed
From Zach Ma and Munehiro Fukuda, “A Multi-Agent Spatial Simulation Library for
Parallelizing Transport Simulations”, WSC 2015

15. Final Remarks Two major agent-based transport simulators: Challenges
CSS434 Demo Talk: Agent-Based Traffic Simulation
Final Remarks
Two major agent-based transport simulators:
TRANSIMS and MATSim (The talk focused on MATSim.)
Challenges
Detailed modeling
Agents and micro-simulation can describe almost whatever we want to model.
Calibrations
Limitation of real data
Labor/data-intensive work
Computational needs
Some parallel/distributed computing efforts have been made.
On-the-fly simulation linked to IoT sensors is not yet addressed because of long-time execution

16. CSS434 Demo Talk: Agent-Based Traffic Simulation
Questions?